Anthropic AI Finds Crispr Like Enzyme Untested
A Discovery Announced Before It Was Tested
In a September 23 announcement, Anthropic claimed that its large language model Claude had identified an enzyme system with properties “reminiscent of Crispr,” the Nobel Prize-winning gene-editing tool. [1] About 950 Claude agents running simultaneously found the genetic sequences in 21.5 hours, according to the company. The announcement arrived before the physical experiments that would validate the finding had even begun. “The experiments are still in the queue. The PR is already live,” says Le Cong, a professor at Stanford University focused on integrating AI into genome engineering research. [1] That gap — between announcement and evidence — is not a footnote to this story. It is the story. A computational pattern match was presented in the language of biological discovery, and the scientific community is now being asked to treat a hypothesis as though it were a result.
What Claude Actually Found
The technical details matter here, because they show how narrow the finding is. Researchers at Anthropic prompted Claude to search genomic databases for “interesting new examples” of reverse transcriptases — proteins that copy RNA into DNA, the opposite of the normal cellular process in which DNA is transcribed into RNA. Claude agents initially identified more than 200,000 possible reverse transcriptases, then narrowed the field to several thousand that appeared new, and eventually surfaced an “unusual” family containing a long region of repeat DNA sequences resembling Crispr. Anthropic calls this system ART, short for array-associated reverse transcriptases. It appears in jumbo phages, large viruses that infect bacteria. One AI agent involved in the search wrote: “I can see by eye a tandem repeat array… that’s a Crispr-like… repeat array?!” The same agent acknowledged the system could be a retron. Crispr and retrons are both bacterial immune systems. Retrons have some utility in gene editing, but they are not the multi-tool that Crispr is. Anthropic’s blog post played up the Crispr comparison anyway.
The Same Enzyme, Found Before
The company itself has been clear the finding is not the end of its work. In a post on X, Anthropic wrote: “We don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA.” [1].” That admission is buried beneath the Crispr framing. The headline says discovery. The fine print says maybe.
A Scientist Who Fed Claude His Unpublished Work
Mario Rodríguez Mestre, a researcher at the University of Copenhagen, had been studying these enzymes. He had not yet published the details. Mestre used Claude extensively during his research process. He has now raised the question of what information the model drew on to arrive at its “discovery.” Anthropic has said the model was not trained on Mestre’s unpublished work, but that denial does not resolve the underlying problem. The scientific community cannot evaluate what data this model was trained on. Major journals require scientists to publish their code freely in the name of reproducibility. Anthropic’s model remains opaque in exactly the way that makes verification impossible. A researcher who had been studying these exact enzymes is left suspicious that the model absorbed his unpublished findings. Whether or not that suspicion is correct, the conditions that produced it are structural.
The Diamond and the Glass
Cong offers the clearest way to understand what has and has not happened. “Let’s say we are on Santa Monica Beach and trying to scan through all the sand to find a diamond,” he says. “AI found this thing that looks very shiny, and then you have to go back to the lab to know — is it glass? Is it a diamond?” [1] The shiny object has been found. The lab work has not been done. The technical report Anthropic released has not been peer-reviewed. It remains unknown whether this enzyme system can be used as a gene-editing tool and, if so, whether it would be a useful one. Identifying new reverse transcriptases can take months of manually mining genome databases. That Anthropic did it in a day is genuinely impressive. But the scientists were not bystanders to their own discovery. The scientists designed the prompts. The scientists chose what to search for. The scientists interpreted the output. The model found patterns in data that humans had already collected and organized. This is not nothing. It is also not autonomous discovery.
The Long Arc from Sequence to Medicine
The history of Crispr offers a corrective to the excitement. Crispr did not become a working tool overnight, and it did not become a medicine overnight either. The distance between spotting a sequence in a database and putting a therapy in a clinic is measured in decades, not hours. That distance is the context in which Anthropic’s 21.5-hour search should be understood. Finding a pattern in a database is the first step of a very long road. The company’s announcement treated it as though it were the destination. Anthropic CEO Dario Amodei mused on X that “Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment.” He quickly followed up by noting that Anthropic’s labs are at the lowest biosafety levels and thus do not contain anything greatly harmful to humans. The idea raises questions about safety and oversight that the company acknowledged but did not engage with. [1].
The Test That Has Not Been Run
Cong proposes a cleaner test for whether an AI model can genuinely discover something in biology. “If you have an AI that only trained on knowledge from before people ever discovered Crispr, and then that AI actually discovered Crispr, that seems to be a better setup to test an AI scientist,” he says. That test has not been run. What has been run is a pattern-matching exercise on data that already existed, using a model whose training data cannot be audited, producing a finding that had already been made by human researchers, announced before the experiments to validate it had begun. The gap between what the AI appears to have done and what it actually did is not a technical failure. The model did what it was asked to do. It found patterns. It surfaced a sequence. It flagged something shiny. The failure is in the telling. A computational result was dressed as a biological discovery. A hypothesis was announced as a finding. A tool that can accelerate the search for patterns was presented as a scientist that can make discoveries. The distinction matters because it shapes what comes next — how resources are allocated, how careers are built, how the public understands what AI can and cannot do. Fyodor Urnov, a gene-editing expert at the University of California, Berkeley and director for therapeutic R&D at its Innovative Genomics Institute, offered genuine praise for Anthropic’s transparency. “I sincerely compliment Anthropic for telling the world about their discovery,” he says. The IGI is collaborating with Anthropic but was not involved with this finding. [1]. That compliment is worth taking seriously. Anthropic did share its results. It did release a technical report. It did acknowledge that the system’s function remains unknown. The company has set up a wet lab for drug discovery and says the work continues. [1].
But sharing a result is not the same as validating it. Releasing a technical report is not the same as peer review. Acknowledging uncertainty in a blog post while leading with a Crispr comparison in the headline is not the same as communicating honestly.
What the Model Cannot Tell Us
The deepest problem with Anthropic’s announcement is not that the finding is wrong. It may turn out to be useful. The problem is that the announcement itself demonstrates the gap between what AI systems can do and what we can verify they have done. Claude found a pattern. We cannot know exactly how. Claude surfaced a sequence. We cannot audit what it was trained on. Claude identified something that looked like Crispr. We cannot know whether the resemblance is functional or superficial without years of laboratory work. The model cannot tell us what it does not know. It cannot distinguish between a diamond and a piece of glass. It cannot tell us whether it learned from published papers or from a researcher’s unpublished chats. It cannot tell us whether the pattern it found is biologically meaningful or merely aesthetically similar to something else. [1]. Those are not limitations of this particular model. They are limitations of the entire approach. An AI system that finds patterns in data is only as good as the data it was trained on, the transparency of its training, and the human judgment that interprets its output. Strip away any of those, and you have a shiny object and a press release. [1]. The science has not caught up with the announcement. The validation has not caught up with the claim. The understanding has not caught up with the discovery. And that is the pattern that matters most — not the one Claude found in the genome, but the one Anthropic followed in telling the world about it.